A Supervised Laplacian Eigenmap Algorithm for Visualization of Multi-label Data: SLE-ML

Abstract

A novel supervised Laplacian eigenmap algorithm is proposed especially aiming at visualization of multi-label data. Supervised Laplacian eigenmap algorithms proposed so far suffer from hardness in the setting of parameters or the lack of the ability of incorporating the label space information into the feature space information. Most of all, they cannot deal with multi-label data. To cope with these difficulties, we consider the neighborhood relationship between two samples both in the feature space and in the label space. As a result, multiple labels are consistently dealt with as the case of single labels. However, the proposed algorithm may produce apparent/fake separability of classes. To mitigate such a bad effect, we recommend to use two values of the parameter at once. The experiments demonstrated the advantages of the proposed method over the compared four algorithms in the visualization quality and understandability, and in the easiness of parameter setting.